WiFi-based human activity recognition (HAR) has demonstrated significant potential in diverse intelligent applications. However, sensitive to environmental factors, cross-domain WiFi channel state information (CSI) poses significant challenges in the generalization of HAR models across different environments. In this paper, the structural causal model (SCM) is introduced to model causal relationships among activities, latent variables, and CSI data, laying a solid foundation toward developing a domain-invariant model for WiFi-based HAR tasks. In this paper, we propose CausalFi, a novel framework that leverages causal inference to mitigate the confounding effects of latent variables, significantly improving generalization performance. Integrating novel feature selection and importance sampling algorithms as the condition and intervention operations, CausalFi can effectively identify the stable action-relevant features from WiFi CSI for activity recognition. Furthermore, a novel counterfactual style augmentation approach is proposed to increase the stylistic diversity of the training data, reducing the risk of biased data distributions even with limited training samples in source domains. We implement a prototype of CausalFi using commercial ASUS RT-AC86U WiFi devices and conduct extensive cross-domain experiments to validate the effectiveness of the proposed approach. With an average recognition accuracy of 92.6%, CausalFi significantly outperforms state-of-the-art baselines in complex cross-domain environments, confirming the practicality of our framework for real-world WiFi-based HAR applications. © 2026 IEEE.